Skip to content
All work

OCR & Document Intelligence

Bangla OCR / ICR Engine

HawkEyes · Bangla-ICR · 2025

End-to-end custom PyTorch intelligent character recognition for the low-resource Bangla script.

PyTorchCNN + CTCOpenCVscikit-learnsynthetic data
CRNN
CNN feature + CTC decode
Bangla
low-resource script
REST
API-served(CV)

Problem

Reading Bangla text is under-served by mainstream OCR engines, blocking automated understanding of Bangla documents and forms.

Approach

  • 01Built an end-to-end PyTorch OCR/ICR model: a CNN feature extractor with sequence decoding under a CTC-style objective and Levenshtein-based evaluation.
  • 02Generated synthetic training data and assembled a custom dataset to overcome the low-resource gap.
  • 03Complemented by a detection-driven field-extraction demo (DemoOCRBangla) and a hybrid Cloud Vision + Gemini OCR pipeline (OCR_CloudVision_Gemeni) for noisy real-world images.

Results & Impact

  • Directly addresses the low-resource-language OCR challenge with a custom, from-scratch CRNN+CTC pipeline and synthetic data.
  • Served as a REST API and paired with hybrid classical-OCR + LLM extraction for messy inputs.— CV-sourced